Maximally Interfered Retrieval plugin

[AlbinSou]

I tried to implement the ER-MIR strategy described in Online Continual Learning with Maximally Interfered Retrieval as a plugin.

The implementation for now is quite slow because of the re-instantiation of the DataLoader at every iteration. I will try to come up with a different solution but right now this is the best-working one that I have so far, so I'll leave this as a basis to improve on.

I also include a replay plugin that works similarly to the MIR one (sample two batches and add the loss coming out). It's just here so that I can compare the two methods fairly.

I thought it would be hard to implement MIR as classical avalanche replay (concatenating replay batch to current batch), since in MIR you have to perform a temporary update with the current batch before knowing what samples to use replay on. So that's why I have to compare to different replay plugin.

I also added a test script with which I try to get as close as the MIR setting described for CIFAR10, that we can use to check the results. Right now I ran only for one seed and had +4% for MIR. But I know a few details are not present in the example script, for instance I don't think they use input transformations on CIFAR10

Feel free to comment and propose improvements. I think the main idea of the method is here but few optimizations can be done still.

[AntonioCarta]

Hi @AlbinSou, thanks for the PR. I need to ask you for two things:

• you have to add a test in tests/training/test_strategies. Just copy-paste the one of the other strategies and adapt it for MIR.
• the example in examples/mir_cifar10 will go in the continual-learning-baselines repository. Can you open a PR there once this is merged?

The implementation looks good to me. Regarding the slow OCL scenario, we are investigating it with @HamedHemati. Hopefully, we will have a solution soon enough.

[AntonioCarta]

Can you describe the difference between the default Avalanche replay and this one here in the docstring?

[HamedHemati]

Thanks, @AlbinSou, for the PR!

@AntonioCarta: One question that I have is: do we want to have a general replay plugin in Avalanche that all replay-based methods can use to implement their strategies?
It, of course, makes sense and would be easier to have strategy-specific internal replay loaders (like the one in this PR) for each new plugin, but I think this plugin can also be implemented without any additional replay loaders after adding the interactive buffer mode to Avalanche ReplayPlugin. It will provide a mask for the samples in each batch so the replay loss can be calculated for buffer samples.

[AntonioCarta]

@HamedHemati good point. A good tradeoff to avoid refactoring the implementation could be to move the "MIRLikeReplay" to the continual-learning-baselines repository as an OCL baseline, while avalanche will keep the new ReplayPlugin. What do you think?

[AlbinSou]

Hey Antonio, yes sure this MIRLikeReplay is just here for the purpose of the comparison so it has nothing to to in avalanche. I will move it to the PR in continual-baseline.

[AlbinSou]

I added the test for now. Just to warn you, I tried to stick closer to the setting by suppressing input rotation and cropping, but this only made things worst .. Now I got 34% for MIR and 37% for replay baselines. I don't know what is wrong. I think for the most part it behaves right (after task 1 2 3 it's above normal replay but last task it drops). Maybe it's just some training artifact that I have to fix. Anyway, I guess later discussions about getting the exact same results than in the paper will happen in the continual-learning-baseline PR.

[AntonioCarta]

Great! as soon as you fix the pep8 errors we can merge the PR. You can check them locally with the command pycodestyle avalanche/ examples/ tests/ --exclude examples/tvdetection

[HamedHemati]

@AntonioCarta Sorry, I missed the reply to my comment. My point was to implement both MIRPlugin and ReplayLikeMIR using the ReplayPlugin from Avalanche. As for MIRLikeReplay, it should be just a script to replicate results for that strategy by setting the correct arguments, and for MIRPlugin, self.replay_loader can be replaced with Avalanche's ReplayPlugin.

[AlbinSou]

@HamedHemati The thing that I had in mind was to implement it like that for now and later change it once the interactive buffer system is on the main branch. I don't know if this answers your question. As for ReplayLikeMIR I think it makes more sense to put it only on the continual-baseline repo since it is not very different from normal replay, I use it more to try and get close to the ER baseline the authors use in their code.

[ContinualAI-bot]

Oh no! It seems there are some PEP8 errors! 😕
Don't worry, you can fix them! 💪
Here's a report about the errors and where you can find them:

avalanche/training/plugins/mir.py:95:81: E501 line too long (82 > 80 characters)
avalanche/training/plugins/mir.py:97:81: E501 line too long (81 > 80 characters)
avalanche/training/plugins/mir.py:173:81: E501 line too long (81 > 80 characters)
3       E501 line too long (82 > 80 characters)

[coveralls]

Pull Request Test Coverage Report for Build 3534234979

• 69 of 72 (95.83%) changed or added relevant lines in 4 files are covered.
• No unchanged relevant lines lost coverage.
• Overall coverage increased (+0.1%) to 73.535%

---

Changes Missing Coverage	Covered Lines	Changed/Added Lines	%
avalanche/training/supervised/strategy_wrappers.py	7	8	87.5%
avalanche/training/plugins/mir.py	54	56	96.43%

Totals
Change from base Build 3521832972:	0.1%
Covered Lines:	13501
Relevant Lines:	18360

---

💛 - Coveralls

[AntonioCarta]

@AlbinSou tests are still failing with this error:

ImportError: cannot import name 'ReplayLikeMIR' from 'avalanche.training.plugins.mir' (/__w/avalanche/avalanche/avalanche/training/plugins/mir.py)

probably because you removed it.

Can you also pull the master and fix the conflicts so that we can merge the PR?

[AntonioCarta]

Thanks @AlbinSou! Please remember to open the PR for the continual-learning-baselines repo.